WebAnything rebuilds real websites into reusable agent environments
Original title: WebAnything: From the Wild Web to Scalable Agent Environments
Fine-tuning Qwen3.5-9B on 5,856 trajectories lifts WebArena-Verified overall success by 13.33 points.
WebAnything is a fully automated framework that reconstructs source-grounded, resettable web environments from real URLs without per-site manual implementation after initial setup and authorization. An Explorer-Builder records browser-observable structures and interactions, and an independent verifier checks frozen candidates and guides local repair or recapture. The authors built 244 website artifacts across 14 domains and 53 categories, supporting 12,984 accepted tasks, and report higher fidelity than human-assisted WebHarbor in a blind audit.
- The set covers 244 artifacts and 12,984 executed, accepted tasks.
- A blind audit gave WebAnything 12 wins, two ties, and no losses against WebHarbor.
- Qwen3.5-9B fine-tuning improved REAL success by 9.9 percentage points.
- Odysseys-Easy Rubric Avg and Task Avg rose by 6.25 and 6.32 points.
- Products
- WebAnythingWebHarbor
- AI models
- Qwen3.5-9B
From the original article
1 min read at huggingface.co
Web agents require diverse, executable, and resettable website environments for reproducible training and evaluation. Human-assisted mirrors demand per-site engineering, while generative construction can diverge from specified real websites. Capture-and-replay preserves observations, with interaction coverage constrained by captured content and replay capabilities.
By Tianci Xie, Zikun Cai, Chaoyun Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.